cloud.verifi/human-verification
Integrates a Telegram bot for human responders (associates), routing verification requests to humans who review claims and answer true, false, or a refined correction via inline buttons.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@cloud.verifi/human-verificationHave a human check whether this customer really opted in to marketing emails."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Verifi
Verified human loops for AI agents. An agent POSTs an intent and a claim; a real human reviews it and answers true, false, or a refined free-text correction. Payment is a native x402 micropayment (USDC or EURC on Base) behind an HTTP 402 paywall: no accounts, no API keys.
Live at verifi.cloud · API docs · llms.txt · MCP endpoint at https://verifi.cloud/mcp
Official MCP Registry name: cloud.verifi/human-verification. Publication and
domain ownership details are documented in docs/MCP_REGISTRY.md.
Why
Every agent framework hits the same wall: the agent cannot trust its own output. Verifi gives the agent a durable human workflow, paid the way agents pay (x402), and designed to remain reliable when a good answer takes several minutes.
Related MCP server: Pingfusi
How it works
agent ──POST /verify──▶ verify-api ──▶ core engine ──▶ Telegram bot ──▶ human
▲ │x402 │ route, │ buttons: │
│ │verify+settle │ audit │ ✅ ❌ ✏️ │
└──────── answer ◀───────┴───────────────┴───────────────┴───────────────┘Three gates, priced in euros: 0.10 EUR admits the request to the human queue. When the human answers within 60 minutes of admission, the result unlocks for 2.90 EUR; a later answer, within 24 hours, unlocks for 1.45 EUR. The price is fixed by when the human answered, never by when the agent unlocks. Pay in EURC, or in USDC converted at the ECB reference rate and rounded up to the cent. The admission asset binds the chain, and every price is in the first
402.No free human work: every chain pays both gates. Human time is the scarce resource and wallet addresses are free to mint, so a free tier would let anyone flood the queue unpaid. Free without a human: the docs, MCP discovery, and the
402response that lets an agent test the connection and read the price.Failure credit: a chain whose admission was actually paid and nobody answers within 24 hours grants one free admission for the next chain, in the same asset. The unlock is not included, and credits cannot be farmed for free work.
Reliable delivery: every call returns
202with a durableverify_id. Agents poll for as long as the human needs, up to 24 hours, or use the optional ready or failed callback.MCP: agents can call
verify_claim,get_verify,unlock_verify, andverifi_infodirectly.
Repository layout
core/ Python core engine
api/ internal FastAPI + admin dashboard + public /contact
bot/ Telegram bot for human responders (associates)
mcp/ MCP server (Streamable HTTP)
routing/ associate selection
payments/ earnings, settlements, payouts
db/ PostgreSQL schema + migrations
instances/
verify-api/ public Verify API (Node.js + @x402/express)
ask-this-finn/ second instance (placeholder)
deploy/ docker-compose, nginx, facilitator config, static site
scripts/ ops wrapper (lock + audit + turn-taking)One VPS runs everything: PostgreSQL (+pgvector), the core engine, the bot, the Verify API, the x402 facilitator, the MCP server, and nginx.
Running it yourself
cp deploy/env.server.example .env # fill in tokens and keys
docker compose -f deploy/docker-compose.yml --env-file .env up -d postgres
docker compose -f deploy/docker-compose.yml --env-file .env up -d --build \
core-api verify-api mcp
docker compose -f deploy/docker-compose.yml --env-file .env --profile bot up -d bot
docker compose -f deploy/docker-compose.yml --env-file .env --profile payments up -d facilitator
docker compose -f deploy/docker-compose.yml --env-file .env --profile edge up -d nginxYou need: a Telegram bot token, a receiving wallet address (X402_PAY_TO), and a small gas wallet for the facilitator (FACILITATOR_PRIVATE_KEY, a few euros of ETH on Base). See docs/WALLETS.md for the money architecture and docs/DECISIONS.md for design decisions.
API in 20 seconds
curl -X POST https://verifi.cloud/verify \
-H "Content-Type: application/json" \
-d '{"intent": "send_email",
"claim": "The user opted in via double opt-in on 2026-07-15.",
"agent_id": "0xYourWalletAddress"}'Every admitted response starts with 202 and status: "processing". Poll until ready or failed; a ready result is retrieved through the separate unlock endpoint. The canonical repository reference is docs/API.md, rendered at verifi.cloud/docs.
License
MIT
This server cannot be deployed
Maintenance
Related MCP Connectors
Ask a real human (not an LLM) for a judgment, $0.10 USDC via x402. The result returns to the agent.
Buy a verified human's approval, verification or judgment at a published price, paid by x402 or card
Human-in-the-loop for AI agents. Submit choices, get a human decision.
Live AI agent trust oracle: audit synthetic IQ, latency & JSON compliance via x402 or free tokens.
Related MCP Servers
- AlicenseNot gradedqualityCmaintenanceEnables AI agents to obtain an independent review of their drafts, returning a pass/fail verdict with specific issues and suggested fixes, plus a signed receipt. Payments are made per check over x402, with no account or API key needed.MIT
- AlicenseNot gradedqualityAmaintenanceEnables AI agents to file review jobs that real human reviewers open, return pinned comments and verdicts on, and approve, so agents can iterate on built sites, videos, and other artifacts based on genuine human feedback.38 npmMIT
- AlicenseNot gradedqualityAmaintenanceEnables AI agents to submit work to real human reviewers, receive pinned feedback and verdicts, and iterate until approved.38 npmMIT
- AlicenseNot gradedqualityAmaintenanceEnables AI agents to call real human reviewers who inspect published work and return pinned comments and a verdict, so the agent can iterate until approved.38 npmMIT